Distributed recommender systems on an opportunistic network environment

Abstract

Mobile devices are common throughout the world, even in counties with limited internet access and even when natural disasters disrupt access to a centralized infrastructure. This access allows for the exchange of information at an incredible pace and across vast distances. However, this wealth of information can frustrate users as they become inundated with irrelevant or unwanted data. Recommender systems help alleviate this burden. The project presents a novel collaborative filtering recommender system based on an opportunistic distributed network. Collaborative filtering algorithms are widely used in many online systems. Often, the computation of these recommender systems is performed on a central server, controlled by the provider, requiring constant internet connection for gathering and computing data. However, in many scenarios, such constraints cannot be guaranteed or may not even be desired. On the proposed recommendation engine, users share information via an opportunistic network independent of a dedicated internet connection. Each node is responsible for gathering information from nearby nodes and calculating its own recommendations. Using a centralized collaborative filtering recommender as a baseline, we evaluate three simulated scenarios composed by different movement speeds and data exchange parameters. Our results show that in a relatively short time, an opportunistic distributed recommender systems can achieve results similar to a traditional centralized system. Furthermore, we noticed that the speed at which the opportunistic recommender system stabilizes depends on several factors including density of the users, movement speed and patterns of the users, and transmission strategies. On future works we will analyze new strategies and datasets, likewise, we will increase the number of users on different scenarios.